Proposal Description

Generalized additive models (GAMs) are among the most widely used tools in applied statistics. Originating from generalized linear models and their popularity, GAMs replace the linear predictor formed from covariates by an additive combination of nonlinear covariate effects. Based on this original model formulation, various extensions have been developed, for example including random effects in generalized additive mixed models to account for unobserved, cluster-specific covariates, regression effects beyond the mean in generalized additive models for location, scale and shape or spatial model variants that account for dependence between observations close to each other in space. Similarly, various modes of statistical inference have been developed, for example utilizing penalized maximum likelihood, Bayesian inference based on Markov chain Monte Carlo simulations or variational approximations, and gradient boosting approaches. As a consequence, GAMs and their extensions have turned into one of the workhorses of statistical modeling.
In this session, we will bring together four experts in GAMs that will present different extensions that are motivated from statistical modeling of complex data. This will include the consideration of novel tools for interpreting, checking and diagnosing GAMs (Fasiolo), challenges arising from the presence of spatial confounding in spatial GAMs (Marques), scalability of Bayesian inferential schemes (Rue), and spatio-temporal modelling of extremes in environmental data (Thorarrinsdottir). With this, the session will provide an overview on current trends in extending GAMs as well as practical orientation for challenges arising from complex data. The presentations in the session will be supplemented by a the contribution of a discussant that adds an integrated perspective on the four presentations and associated points for discussion (Groll).